Work

Physical Fidelity of Learned Nonlinear Vehicle Dynamics: Validation Selection, Observational Bias, and Architecture-Conditioned Inductive Bias

Year: 2026
Type: preprint
Abstract: Learned vehicle-dynamics models can reproduce nominal trajectories while misrepresenting local tangent and finite-time behavior. We evaluate this discrepancy using analytical 3-, 7-, and 14-state nonlinear vehicle models with vector-field, Jacobian, directional-deformation, and frozen-local propagator metrics. Reference-based validation selectors substantially enrich critical-region coverage: at a 10\% budget, symmetric growth reaches macro recall 0.4338 versus 0.1011 for random sampling. In fresh-seed training, physics-inspired selection reduces 14-state Jacobian error by 2.09\%, below the predefined 5\% practical threshold, while the 5.80\% propagator reduction is not significant after Holm correction. A coverage-constrained Lagrangian selector improves Jacobian error by 16.5\% but worsens propagator error by 27.6\%, showing that better local derivative fidelity does not guarantee better finite-time response. Adaptive loss balancing and Adam-to-L-BFGS refinement provide no consistent gains, and architecture effects depend on learner parameterization. A post-hoc state-resolved analysis of 716 frozen checkpoints further shows that training interventions redistribute fidelity across physical states: M1 improves the longitudinal, pitch-rate, and wheel-spin equations while degrading lateral-velocity and roll-rate dynamics. These results support target-specific, coverage-aware training and validation rather than assuming that more physics-selected data or stronger physical priors improve all fidelity measures.
Language: English
Open Access status: green
OpenAlex ID: W7233607033
Cites:
Cited by:
Related to: